TL;DR
- Live data improves a model only when it is aggregated per cell and reconciled with the site database; raw dumps make predictions worse.
- The highest-value inputs are timing advance distributions, per-cell PRB utilisation, and measurement-report RSRP and RSRQ distributions.
- Quarterly recalibration beats continuous recalibration: models drift slower than teams assume, and continuous tuning fits noise.
Which live data sources actually improve a prediction?
Three sources carry most of the value. Per-cell timing advance distributions give an independent estimate of how far users actually are from the antenna, which is the single best calibration signal for path loss exponent and clutter offset. Per-cell physical resource block utilisation shows where the model predicts capacity that does not exist. Measurement-report based RSRP and RSRQ distributions give the received-level distribution the model should reproduce.
Everything else is secondary. Handover counts help validate neighbour predictions, and per-cell throughput helps validate the capacity model, but neither corrects a path loss error. The common mistake is to import every counter available because the OSS exports them, then discover that the model has been tuned against a mixture of metrics with different biases.
| Data source | What it corrects | Typical pitfall |
|---|---|---|
| Timing advance distribution | Path loss slope and cell radius | Aggregated over all technologies, mixing RAT-specific timing granularity |
| MR-based RSRP/RSRQ | Received level distribution per cell | Reported only by connected users, so it is biased toward the cell edge |
| PRB utilisation | Capacity and load distribution | Busy-hour definition differs between OSS and planning assumptions |
| Handover counters | Neighbour overlap | Includes failed handovers, which are exactly the ones the model got wrong |
| Drive test logs | Absolute level and quality truth | Small sample, biased to roads, no indoor coverage |
How do you reconcile OSS counters with the site database?
Reconciliation is the step that decides whether the data helps or hurts. Every counter record must map to a cell identity that exists in the planning tool with the same azimuth, tilt, height and antenna model. In multi-vendor networks the same physical sector can appear under three different naming conventions, and a single mismatched cell will drag the calibration in the wrong direction.
A practical approach is to build a mapping table once, keyed on the cell identity used by the OSS, with explicit columns for planning-tool cell name, site ID, sector, and the parameter values as actually installed. Then audit it: any cell in the mapping table whose installed tilt differs from the planned tilt is a data error that will corrupt the model if it is not fixed first.
What does a recalibration workflow look like?
Collect a defined window — typically two to four weeks of busy hours, avoiding festivals and network events. Extract per-cell aggregates, not raw samples. Compare predicted and measured distributions per clutter class rather than per cell, because per-cell comparison over-fits to local measurement noise.
Then adjust the model in the smallest number of places that explains the largest error. If every suburban cell is 5 dB optimistic, correct the suburban clutter offset; do not add per-cell correction factors. Per-cell corrections produce a model that reproduces last month perfectly and predicts nothing.
When does live data make a model worse?
Three situations. First, when the measurement sample per cell is too small: a rural cell with a few hundred measurement reports in a month produces a distribution dominated by where the few users happened to be. Second, when the network is not in the state the model represents — during a rollout, half the cells are not yet carrying traffic at design load. Third, when the data is used to correct a structural error: if the site database has the wrong antenna height, calibration will compensate with a clutter offset that is wrong everywhere else.
The test is simple. Calibrate on one cluster, then predict a second cluster that was not used in calibration. If the error does not improve on the held-out cluster, the calibration fitted noise.